Parallelization of SAT Algorithms on GPUs [pdf]

This approach is unworkable, it has been demonstrated many times. I have not only been developing a SAT solver for >7 years (with >9000 commits) but also have participated in the SAT competition in all these years. I used to get a proposal like this a lot (thankfully, they dried up). None of them made it work. I used to work on programming GPGPUs in a professional setting, (rainbow table generators&lookup systems), so I'm not unfamiliar with their advantages and restrictions.

This approach is unworkable because: * propagation needs modification of the watchlist (or it needs to have all literals of all clauses attached, which is incredibly slow and also unworkable) * propagation is actually only about 50-60% of what a solver is doing: conflict resolution, clause database management and inprocessing takes a lot of time * propagation is hard to parallelize as you need to update a global data structure of propagated literals and then act on that * etc

This proposal will not work (and I sure it did not, if it was ever tried) and I don't understand how it got this high at ycombinator. If you want to look at good SAT solvers, look at lingeling, COPMinisat, MiniSat, or CryptoMiniSat. You may learn a lot from them. Easy starting is MiniSat. Good luck :)

There is quite a bit of work on parallelising SAT solving, as it's one of the most important algorithmic problems, both in theory and in practise, but (as far as I'm aware) the top sequential algorithms still beat parallel solvers. The most successful approach at this point seems to be Cube-and-Conquer [1], but I don't think this is considered a good approach in itself. There is a huge amount of scope for new ideas. I doubt that anything naive will work, as SAT solving is a rather competitive research field.

The annual SAT competition http://www.satcompetition.org has a track for parallel solving.

See [2, 3, 4] for an overview of the field.

[1] M. J. H. Heule, O. Kullmann, S. Wieringa, A. Biere, Cube and Conquer: Guiding CDCL SAT Solvers by Lookaheads.

[2] S. Hölldobler, N. Manthey, V. H. Nguyen, J. Stecklina, P. Steinke, A Short Overview on Modern Parallel SAT-Solvers.

[3] S. Wieringa, K. Heljanko, Asynchronous Multi-Core Incremental SAT Solving.

[4] Y. Hamadi, C. M. Wintersteiger, Seven Challenges in Parallel SAT Solving.

SAT solvers and neural networks are probably the most curious software technologies around. Both attempt to solve NP-complete problems (learning and SAT). Both are still largely based on 1960s algorithms (backpropagation for NNs and DPLL for SAT) made surprisingly useful in practice by advances in hardware and heuristics accumulated over the years. Both are mysterious from a theoretical point of view, with little progress in our theoretical understanding of why they work well in practice and how we can come up with better algorithms.

This article by Moshe Vardi[1] and this one by Dick Lipton[2] are a nice introduction to the mystery of SAT solvers.

[1]: http://cacm.acm.org/magazines/2014/3/172516-boolean-satisfia...

[2]: https://rjlipton.wordpress.com/2009/07/13/sat-solvers-is-sat...

This is a proposed work for a Master's thesis. It seems to be from 2013, so I wonder whether it was successful?

If I had to guess, I would think implementing a SAT solver on a GPU is quite hard. Many branches etc.

Publish or perish leaves very little room for sanity checks once your window of graduation starts to close.